plate I.I · the language

i / iv

cr8script·atlas

an atlas of a small language, drawn for the agents who write it

cr8script is an English-shaped scripting language for LLMs and anyone who wants quick scripts without Python's footguns. One file. No build. No imports. It refuses truthy/falsy, refuses silent type mixing, and prints 0.1 + 0.2 = 0.3 exactly — decimal-safe by default, useful for finance, billing, and reporting scripts where Python's float footguns would silently corrupt totals. Pipelines on lists of records read like SQL; errors name the line, the value, and the fix.

view on github ↗ see a sample ↓ llms.txt ↗

chart key Syntax Pipelines Types Built-ins Tooling Principles

plate I.II · the atlas

ii / iv

fig. i -- the language, charted by category generated by examples/make_mindmap.cr8

plate II.I · what makes it different

iii / iv

The premise is unfashionable: a language whose refusals are the feature. cr8script will not let "5" + 3 be a string. It will not let if 0 then be falsy. It will not silently swallow a typo'd field name. The error messages know the line and the value, and offer a fix.

refused, by design

truthy / falsy * == != >= <= * silent type coercion * null / undefined * classes & methods

kept, with intent

let & var * nothing * is at least * for each in * group by * summarize * f"{x}"

made for agents

A surface so small that an LLM rarely gets it wrong. The static checker emits { line, message, hint } JSON so a model can self-correct before running.

plate II.II · a taste

iii / iv

a taste, in cr8

fig. ii -- pipelines · group by · summarize

sample.cr8# sales -- list of records
let sales = [
  { product: "widget", region: "east", amount: 12.50 },
  { product: "gadget", region: "east", amount:  8.00 },
  { product: "widget", region: "west", amount: 15.00 },
  { product: "doodad", region: "east", amount: 99.00 },
]

let by_product = sales
  | group by product
  | summarize { total: sum(amount), n: length(items) }
  | sort by total descending

for each row in by_product
  show f"{row.product}: {row.total} across {row.n} order(s)"
end

stdoutdoodad: 99 across 1 order(s)
widget: 27.5 across 2 order(s)
gadget: 8 across 1 order(s)

plate III.I · six provinces

iv / iv

the six provinces

fig. iii -- the whole surface area

i. syntax

Syntax

let immutable · var mutable
to name(...) defines a function · end closes the block
if · then · else · end -- also an expression
for each n in 1..5
repeat 3 times
try / otherwise as err
f"hello {name}" interpolation

ii. pipelines

Pipelines

where · filter
sort by · ascending / descending
take · first n
map · transform
group by · yields {key, items}
summarize { ... } · roll up
bare names auto-resolve to record fields

iii. types

Types

number -- one decimal type, 0.1+0.2=0.3
text -- sealed strings, no coercion
boolean -- only in conditions, no truthy/falsy
list -- 1-based indexing
record -- values, not objects
nothing -- the only absence

iv. built-ins

Built-ins

math · sqrt · pow · pi · e
http · get returns { ok, status, body }
time · now · sleep
json · parse · stringify
csv · parse · write
sum / count / average / min / max

v. tooling

Tooling

pip install cr8script · one command on PATH
--check-json · structured diagnostics for agents
--test · golden suite
tour.cr8 · the language end-to-end
AGENTS.md · system-prompt template + tool defs
LLM_MAP.md · typed planning graph

vi. principles

Principles

One way to do each thing
No invisible syntax -- blocks end with end
Reads like English
Strong, honest types
Errors that teach (line · value · hint)
Determinism over cleverness

plate IV.I · run it

v / vi

run it

python 3.9+ · one file · no dependencies

$ pip install cr8script install

$ cr8script --example tour language tour

$ cr8script --help all flags

$ cr8script --check-json file.cr8 llm correction loop

plate IV.II · for language models

vi / vi

for language models

discovery · llmstxt.org · self-correction

cr8script is built so an agent can write it correctly the first time. Five surfaces exist for that purpose:

/llms.txt -- discovery file at this site's root, following the llmstxt.org convention. A pointer table, in markdown.
LLMS.md -- a condensed grammar & rules sheet inside the language repo. Hard rules first, then the cheat sheet, then idioms.
AGENTS.md -- the operational playbook: a paste-able system-prompt block, tool definitions for cr8_check / cr8_run, and a task-shape pointer table. Everything you need to wire cr8script into an agent loop.
LLM_MAP.md -- a typed planning graph that sits between prompt and .cr8 generation. Forces the model to externalize goals, inputs, transforms, decisions, risks, outputs, and checks before writing code. Has its own validator (tools/check_map.py) that mirrors the --check-json shape.
--check-json emits { line, message, hint } diagnostics -- iterate against the static checker before running.
Longer-form pages on the same topic: principles (the design defended), cookbook (recipes for common tasks), and cr8 vs python (four scenarios where Python is silent and cr8 is loud).